Andes Technology Corporation (TWSE: 6533; SIN: US03420C2089; ISIN: US03420C1099), a leading supplier of high efficiency, low-power 32/64-bit RISC-V processor cores and Founding Premier member of RISC-V International, submits MLPerf Tiny v0.7 benchmark for Andes V5 RISC-V processors, including the AndesCore™ D25F and D45 (with DSP/SIMD Extension) and AndesCore™ NX27V (with Vector Extension). The MLPerf Tiny benchmark suite is intended for the lowest power devices and smallest form factors, such as deeply embedded, intelligent sensing, and internet-of-things applications. The remarkable out-of-box results reveal the outstanding Andes processor capabilities with performance and power consumption.
The out-of-box results of the MLPerf Tiny v0.7 are enabled by high-performance AndesCore™ and the enhanced “TensorFlow Lite for Microcontroller” with “Andes NN Library”. AndesCore™ V5 processors with DSP/SIMD (P draft) and Vector (V) Extensions are designed for applications needing highly efficient or intensive data computation. RISC-V DSP/SIMD Extension (RVP) efficiently addresses the requirements of low-volume data computation with straightforward hardware implementation for low power consumption without handcraft optimization in the model and software package. According to the Visual Wake Words and Image Classification scores, at the same 200MHz, Andes V5 RVP cores can deliver 11 fps on the D25F and 15 fps on the D45F while typical MCUs are only capable of 6-7 fps. By providing the compact and efficient SIMD and DSP capabilities, the D25F and the D45 processors form a very comprehensive and competitive basis for the TinyML, AIoT, and signal processing applications on edges and endpoints. RISC-V Vector Extension (RVV) targets high-volume data computation. No matter in the edge or cloud, the NX27V processor provides scalable, efficient, and powerful compute capabilities for general AI, NN, and data processing applications. Moreover, “Andes NN Library” dramatically speeds up the development of Neural Network algorithms with 17.3x speedup of MobileNet-v1 INT8 using AndesCore™ NX27V with the 128-bit vector configuration over the same core executing only RISC-V baseline (scalar) instructions.
“AndesCore™ RISC-V processors make intelligence available for TinyML or MCU AI without needing extra cost for hardwired accelerators. It also could be part of an efficient AI subsystem along with an accelerator to offload unstructured computation or pre-/post-processing. It brings unique and competitive value to the customers,” said Dr. Charlie Su, Andes Technology President and CTO. “AndesCore™ processors with RVP and RVV offer comprehensive solutions to the demand for computing in every aspect. In addition, with the optimized compute libraries and tools for RVP and RVV, customers could accelerate the development with highly competitive solutions to facilitate AI applications for a wide variety of products.”